[datlechin]

Summary

• scripts/benchmark-large-catalog.sh times, on the Measure the sidebar, Open Quickly and autocomplete on 100 schemas and 50,000 tables #3661 fixtures (PostgreSQL and Oracle with 100 schemas of 500 tables, MySQL with one database of 50,000), the sidebar's first listing after connect, one filter keystroke and opening a schema, Open Quickly's panel and first result, completion after FROM and after schema. (MySQL: database.), and resident memory. 10 runs each, budgets in scripts/benchmark/large-catalog/baseline.json (1.25x the first run's median and p95). A step also fails when the main thread stalls for more than 100 ms during it.
• The app times these itself with TABLEPRO_INTERACTION_TRACE=1: InteractionTracer (Interaction intervals under Points of Interest, one stderr line each) and MainThreadStallMonitor (each main run loop turn of 50 ms or more). New CodeSuggestionDelegate.completionWindowDidShow(textView:itemCount:).
• Sidebar: a wake from a shared service rebuilds the outline only when what this connection draws moved (DatabaseTreeRenderInputs); filter buckets, listing matches and table rows are memoized across refreshes; SidebarSearch ranks in one pass; the listing's 50,000-row mapping runs off the main actor.
• Sidebar search on a one-row-per-schema tree (Oracle, Dameng, SAP HANA, Trino) draws a schema it opens from the all-schema listing instead of reading it, and reads no schema until that listing lands: a one-character search queued 19 schema reads ahead of it on the one metadata connection, and the first keystroke took 8.1 s.
• Oracle schema rows are selectable, so the keyboard can reach and open them.
• Oracle reads that keep every row fetch 1,000 rows per round trip: the 50,000-table listing went from 1,001 fetches to 51.
• Open Quickly builds its table rows off the main actor (each open rebuilt 50,000 rows on it), rebuilds them only when the rows it reads change, and stops a superseded ranking early.
• Autocomplete ranks off the main actor, keeps the pool for the prefix it answered, and asks again when a deleted character leaves a prefix that pool cannot answer. A pool cut at 400 no longer hides the best match. schema. finds an unquoted schema in any case and keeps a quoted one as written. database. lists that database's tables on an engine without schemas.
• AI chat ranks the schema off the main actor when a message is sent. The footer @ Tables menu lists the first 100 tables; @ in the message finds any other.
• MCP list_indexes reads a whole schema's indexes in one query where the driver supports it.
• Menu validation probes a driver's statement hooks once per driver.
• Fixture script: MySQL over a socket, sessions start in the first schema.

Before is the first run (2026-10-22, now the baseline), after is this branch. Release build, 10 runs, Mac16,11 (M4 Pro, 24 GB, macOS 27.0.1), PostgreSQL 17 local, Oracle Free 23 and MySQL 8.4 in containers, load 4 to 8.

Fixture	Step	Before ms (median / p95)	Before worst stall	After ms (median / p95)	After worst stall	Budget
PostgreSQL	sidebar-listing	445 / 506	271	336 / 449	251	560 / 635
	sidebar-filter	124 / 380	192	89 / 472	70	155 / 480
	sidebar-expand	133 / 243	58	79 / 112	0	170 / 305
	quick-open-panel	170 / 261	51	62 / 98	91	215 / 330
	quick-open-result	236 / 369	0	174 / 183	0	295 / 465
	complete-from	85 / 92	0	69 / 72	0	110 / 120
	complete-schema	122 / 140	0	75 / 83	0	155 / 175
Oracle	sidebar-listing	1,445 / 2,185	964	1,311 / 2,077	787	1,810 / 2,735
	sidebar-filter	138 / 668	195	103 / 1,046	82	175 / 840
	sidebar-expand	not measurable		282 / 338	0	355 / 425
	quick-open-panel	173 / 203	60	77 / 93	86	220 / 255
	quick-open-result	172 / 208	0	156 / 168	0	215 / 260
	complete-from	84 / 91	0	68 / 75	0	105 / 115
	complete-schema	499 / 540	0	442 / 508	0	625 / 675
MySQL	sidebar-listing	574 / 658	222	476 / 544	217	720 / 825
	sidebar-filter	134 / 196	140	81 / 84	71	170 / 245
	sidebar-expand	286 / 317	112	244 / 251	96	360 / 400
	quick-open-panel	156 / 186	0	73 / 91	78	195 / 235
	quick-open-result	142 / 150	0	126 / 138	0	180 / 190
	complete-from	103 / 110	0	77 / 87	0	130 / 140
	complete-schema	completed nothing		95 / 116	0	120 / 145

Resident memory with the table list loaded: PostgreSQL 182 MB, Oracle 162 MB, MySQL 492 MB (before: 119, 194, 772).

Still over:

• sidebar-listing stalls 200 to 310 ms on every fixture, the same on a two-table PostgreSQL database: it is the connection window's first SwiftUI build (editor, results pane, sidebar rows, toolbar), not the catalog. It needs its own design.
• Oracle sidebar-filter p95 is the first keystroke of the session, which waits for the all-schema listing; Oracle's own time for that query on this fixture was 0.55 to 1.2 s (V$SQL). The other nine runs were 80 to 109 ms.

Before / After

Oracle, typing LC_S050 in the sidebar. Before: type-select skips schema rows and lands on a table. After: the schema row is selected.

	Before	After
Light
Dark

AI chat footer @ > Tables on a schema of 500 tables. Before: all 500. After: the first 100 and a hint to type @.

	Before	After
Light
Dark

Test plan

• New: DatabaseTreeReconcileTests, DatabaseTreeFilterMemoTests, DatabaseTreeHierarchicalSearchTests, InteractionTraceRecorderTests, DatabaseObjectToolSupportTests, AIChatTableMentionsTests, AIChatViewModelSchemaRankingTests, MCPListIndexesTests, OracleRowLimitsTests, and cases in QuickSwitcherListingDemandTests, SQLSchemaProviderUnqualifiedScopeTests, SQLCompletionProviderCatalogTests, SQLContextAnalyzerQuotedReferenceTests, DatabaseTreeNodeTests, SuggestionLiveCursorTests, SidebarSearchTests, SidebarNameFilterTests.
• Changed on purpose: AllSchemaTablesDemandTests.resetForgetsEarlierRequests (a reset asks again only for a listing that is not current).
• Local: 1,027 hosted and package-in-plan tests across the sidebar, Quick Switcher, autocomplete, AI chat and editor suites; swift test for TableProAutocomplete (475) and TableProOracle (316); lint clean.
• Keys in the benchmark came from a library loaded into the measured copy that posts CGEvent-built events into its own queue: this session has no Accessibility grant, and XCTest's Automation Mode asks for the owner's password.

Refs #3661